Automations with one AI step
Most “AI automation” is ordinary automation in which AI does one thing: classifies a ticket, extracts data from a PDF, writes a summary. The rest is rules you can check. That is how I build it.
What I implement
Ticket classification
An email, a form submission or a chat message gets a category, an urgency and an owner. Uncertain cases go to a queue for a person, not to a random department.
Data from documents
Invoices, orders, contracts, CVs: from a PDF or scan into fields in your system. With a confidence level on every field and a queue for manual checking.
Summaries and notes
A meeting recording, a phone call, an email thread: a note in the CRM with what was agreed and the next step. Approved with one click.
Content from your data
Product descriptions from a spreadsheet of attributes, language versions, draft posts. In your company’s voice. A person approves publication.
Matching and tidying data
Merging duplicates, matching payments to invoices, filling missing fields from documents. AI proposes, a rule checks, a person approves the exceptions.
Platforms
Make, n8n, Zapier, Power Automate, Workato. When there are several connections, the logic changes, and someone on your side wants to edit it.
Custom code
When fixed cost matters at high volume, when there is no connector, or when the logic is too much for building blocks. With a log, retries and alerts.
Inside Salesforce
Flow and Prompt Builder when everything happens in Salesforce. One AI step in an ordinary Flow, on a separate user.
How it goes
A map of the workflow
What comes in, what goes out, who does it by hand today, what happens on error. An hour’s conversation and a few real examples.
One AI step
Exactly which step AI does, and how we will know it did it right. The rest of the workflow is rules.
A prototype on past data
The last hundred tickets or invoices. AI’s result next to the person’s decision. You know the agreement rate before anything goes live.
Go-live with an exceptions queue
Confident cases flow through on their own; uncertain ones wait for a person. We set the confidence threshold together and can move it.
Care
New document types, a supplier changing its format, a new model. A monthly review of the exceptions queue: what AI gets wrong most often.
Platform or custom code
A platform (Make, n8n, Zapier, Power Automate, Workato)
Quick start, ready connectors, changes without a developer. Cost grows with the number of operations. Good when there will be several workflows and someone on your side wants to develop them. I document so that you can.
Custom code
A fixed cost regardless of the number of documents, full control over logic and data, easier testing. Needs someone to maintain it. Good for high volumes, unusual logic, and data that should not pass through yet another vendor.
I often combine the two: a platform to connect the systems, custom code for the one hard step.
Questions about automations
Make, n8n or custom code?
It depends on the number of workflows, the volume, and who will develop them. If you already have one of the platforms, I start there. You get a proposal with pros and cons after the first call.
How accurate is data extraction from PDFs?
It depends on the documents: clean invoices from regular suppliers come out very well, scans and unusual layouts less so. That is why I measure it on your own sample before go-live, and fields with low confidence go to a person, not into the system.
Can AI post invoices to the ledger on its own?
I do not propose that at the start. AI extracts the data and proposes the coding; a person approves. After a month, for regular suppliers and amounts below a threshold you set, some cases can pass through on their own. The rules are in the article on the human in the loop.
What does it cost to run?
Three components: the model cost per document or ticket (which I show up front, based on the prototype), the platform subscription if we use one, and my hours for care. Quoted after a call, depending on scope.
We already have Zapier and a few automations.
Good. I review them, add an AI step where a person makes a manual decision today, and tidy what has sprawled. I do not rebuild from scratch.
Is there a task someone at your company does by hand every week? Describe it in one sentence.
I reply within one business day. The first call is free of obligation.
Describe what you want automated
One sentence about the task is enough to start. You do not need to know the names of any tools. I reply within one business day.